A general Neural Particle Method for hydrodynamics modeling

Neural Particle Method (NPM) is a newly proposed Physics-Informed Network (PINN) based, truly meshfree method for hydrodynamics modeling. In the NPM, PINN applied to globally approximate field variables, and high-order Implicit Runge–Kutta (IRK) used treat time integration. The NPM can easily achieve incompressibility deal with free-surface problem. However, currently limited inviscid problems computationally expensive. this work, we developed general (gNPM) viscous gNPM, single pressure output as predicted rather than multiple pressures in original NPM. Thus, size of neural network greatly reduced, making gNPM more efficient. Besides, spatial derivatives governing equations are calculated respect current coordinates space. manner, straightforward be implemented. Furthermore, by considering term conservation momentum, effectiveness robustness have been demonstrated through several benchmark cases different boundary conditions. We highlight that able cope highly uneven particle distributions, while traditional produce severe failure.